Region-wise comparison assessment on PANDABENCH (Hard set)
43AccuracyLinear Probe
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Linear ProbeType=Baseline2026.04 | 43 | 28 | 21 | 14 | |
| PANDA2026.04 | 40 | 31 | 25 | 24 | |
| Attentive ProbeType=Baseline2026.04 | 39 | 23 | 21 | 18 | |
| DepictQA†Access=Open-source, Training=Trained on PANDASET2026.04 | 33 | 21 | 18 | 19 | |
| GPT-5 NanoAccess=Closed-source2026.04 | 21 | 17 | 16 | 16 | |
| RandomType=Baseline2026.04 | 21 | 20 | 20 | 19 | |
| GPT-5 MiniAccess=Closed-source2026.04 | 18 | 19 | 20 | 15 | |
| GPT-4oAccess=Closed-source2026.04 | 16 | 19 | 16 | 14 | |
| Gemini 2.5 ProAccess=Closed-source2026.04 | 12 | 20 | 14 | 12 |